Spectrum MiniMax H3 v0.2.1

Spectrum MiniMax H3 v0.2.1

Makes the validated H3 audio-quality correction the standard default path.

Corrected default

• offline_smoothing_replay now defaults to true in the runtime configuration, node schema, and Python call signature.
• The mode is no longer classified as experimental.
• The default remains blend_weight=0.5 for video and audio_blend_weight=0.0 for audio.
• Offline capture uses causal video=0, audio=0; configured per-modality weights are applied only during transformer-free replay.

This default follows the matched full-checkpoint result from v0.2.0 development. Single-pass video=0.5, audio=0 reproduced degraded speech and stuttering because video forecasts fed into later joint audio-video transformer calls. Offline capture/replay with the same weights removed the defect and restored high-quality audio on that seed while retaining the preferred image result. Disabling offline replay brought the defect back.

Retained research modes

• anchor_residual_feedback remains experimental and default-off.
• selective_rollback_correction remains experimental and default-off.
• Both remain mutually exclusive with each other and with the standard offline path. Disable offline_smoothing_replay before enabling either research mode.
• Explicitly setting all three trajectory modes to false retains the single-pass comparison path.

Upgrade behavior

New nodes use offline replay automatically. Workflows created before v0.2.0 did not store this input and therefore also receive the new default. Workflows saved with v0.2.0 may retain a serialized offline_smoothing_replay=false; enable it once in those existing nodes to use the corrected path.

Offline replay retains every actual hidden anchor plus cloned sampling inputs. history_storage=system_ram remains the default; vram avoids large host transfers when enough VRAM is available. Unsupported samplers run one valid native pass, and an incomplete replay archive returns the valid local-only first-pass result. Cancellation, OOM, and other fatal exceptions preserve normal teardown and propagation.

Validation

• Default-value tests cover the dataclass, ComfyUI node schema, and Python call signature.
• The outer-sampler integration test now proves that the omitted/default setting executes capture plus transformer-free replay.
• Existing single-pass, residual-feedback, rollback, native-equivalence, t2va, fl2va, and ref2va contracts remain covered.